Training Samples Research Articles

Prediction of acetylene solubility by a mechanism-data hybrid-driven machine learning model constructed based on COSMO-RS theory

In this work, a mechanism-data hybrid-driven machine learning model is built to predict acetylene solubility from COSMO-derived molecular descriptors, overcoming the challenge of limited training samples. We have successfully enhanced the data-learning capabilities of the model by mechanism-level modelling, and generated new mechanism-level understanding from the results of data learning. On the basis of the mechanism-level understanding of molecular interactions, this model gives more accurate predictions than the baseline models, even though it contains only 37 model parameters. Meanwhile, it presents better generalization to predict solubility in solvents beyond the training sample. Furthermore, by analyzing the implication of the parameters of this model, we find a charged segment interaction energy form different from that described by the classic COSMO-RS theory. This work not only produces available acetylene solubility data, but also presents a proposed framework to build a mechanism-data hybrid-driven model for thermodynamic properties and discovering interpretable chemical laws from small training samples.

AIPs-DeepEnC-GA: Predicting Anti-inflammatory Peptides using Embedded Evolutionary and Sequential Feature Integration with Genetic Algorithm based Deep Ensemble Model

Inflammation is a biological response to harmful stimuli including infections, damaged cells, tissue injuries, and toxic chemicals. It plays an essential role in facilitating tissue repair by eliminating pathogenic microorganisms. Currently, numerous therapies are applied to treat autoimmune and inflammatory diseases. However, these conventional anti-inflammatory medications are often labor-intensive, costly, and associated with adverse side effects. Recently, researchers have identified anti-inflammatory peptides (AIPs) as a cost-effective alternative for treating several inflammatory diseases, due to their high selectivity for target cells with minimal side effects. In this study, we introduce a novel computational predictor, AIPs-DeepEnC-GA, developed to accurately predict AIP samples. The training sequences are encoded using a novel n-spaced dipeptide-based position-specific scoring matrix (NsDP-PSSM) and Pseudo position-specific scoring matrix (PsePSSM)-based embedded evolutionary features. Additionally, the reduced- amino acid alphabet (RAAA-11), and composite Physiochemical properties (CPP) are employed to capture cluster-physiochemical properties based on structural information. A hybrid feature strategy is then applied, integrating embedded evolutionary features, CPP and RAAA-11 descriptors to overcome the limitations of individual encoding methods. Minimum redundancy and maximum relevance (mRMR) is utilized to select the optimal features. The selected features are trained using four different deep-learning models. The predictive labels generated by these models are provided to a genetic algorithm to form a deep-ensemble training model. The proposed AIPs-DeepEnC-GA model achieved a ∼15% increase in predictive accuracy, reaching 94.39%, and a 19% improvement in the area under the curve (AUC), achieving a value of 0.98 using training sequences. For independent datasets, our method obtained improved accuracies of 91.87%, and 89.21%, with AUC values of 0.94 and 0.92 for Ind-I, and Ind-II, respectively. Our proposed AIPs-DeepEnC-GA model demonstrates an 11% improvement in predictive accuracy over existing AIPs computational models using training samples. The efficacy and reliability of this model make it a promising tool for both in drug development and research academia.

A fast deployable model for crack identification with laser thermography testing

This paper presents a novel feature extraction method, enabling efficient training and deployment of neural networks for rapid identification of crack defects in Laser Array Spot Thermography (LAST). We trained the crack defect identification model based on pixel-level features, encoding each pixel as a feature vector using Frangi filter, and classifying them using a neural network. Experimental results demonstrate that Frangi features are an effective method for distinguishing cracks, speckles, and background noise interference in the experiment. Furthermore, the model only requires a small region of interest (ROI) as training samples to achieve effective training and efficient crack identification under the same detection conditions, allowing for rapid deployment in practical inspections.

Plug-and-play segment anything model improves nnUNet performance.

The automatic segmentation of medical images has widespread applications in modern clinical workflows. The Segment Anything Model (SAM), a recent development of foundational models in computer vision, has become a universal tool for image segmentation without the need for specific domain training. However, SAM's reliance on prompts necessitates human-computer interaction during the inference process. Its performance on specific domains can also be limited without additional adaptation. In contrast, traditional models like nnUNet are designed to perform segmentation tasks automatically during inference and can work well for each specific domain, but they require extensive training on domain-specificdatasets. To leverage the advantages of both foundational and domain-specific models and achieve fully automated segmentation with limited training samples, we propose nnSAM, which combines the robust feature extraction capabilities of SAM with the automatic configuration abilities of nnUNet to enhance the accuracy and robustness of medical image segmentation on smalldatasets. We propose the nnSAM model for small sample medical image segmentation. We made optimizations for this goal via two main approaches: first, we integrated the feature extraction capabilities of SAM with the automatic configuration advantages of nnUNet, which enables robust feature extraction and domain-specific adaptation on small datasets. Second, during the training process, we designed a boundary shape supervision loss based on level set functions and curvature calculations, enabling the model to learn anatomical shape priors from limited annotationdata. We conducted quantitative and qualitative assessments on the performance of our proposed method on four segmentation tasks: brain white matter, liver, lung, and heart segmentation. Our method achieved the best performance across all tasks. Specifically, in brain white matter segmentation using 20 training samples, nnSAM achieved the highest DICE score of 82.77 ( 10.12) % and the lowest average surface distance (ASD) of 1.14 ( 1.03) mm, compared to nnUNet, which had a DICE score of 79.25 ( 17.24) % and an ASD of 1.36 ( 1.63) mm. A sample size study shows that the advantage of nnSAM becomes more prominent under fewer trainingsamples. A comprehensive evaluation of multiple small-sample segmentation tasks demonstrates significant improvements in segmentation performance by nnSAM, highlighting the vast potential of small-samplelearning.

ECGWiz: a unified and generalizable AI model for comprehensive ECG diagnosis

Abstract Background Deep learning has demonstrated diagnostic performance in electrocardiogram (ECG) analysis comparable to that of clinicians. However, existing models primarily excel in diagnosing common conditions such as arrhythmias and heart blocks, which are easily identifiable using standard predictor variables. Conversely, models aiming to uncover novel insights, such as detecting structural changes or the presence of myocardial scar, often lack generalizability due to insufficient data, model architecture limitations, and other factors. Moreover, these models frequently encounter challenges, such as a decline in performance when applied to external datasets. To address these issues comprehensively, we propose the ECGWiz — a unified model capable of tackling diverse ECG-related problems with consistently high performance. Objective To develop a generalizable Foundation AI model for ECG diagnosis capable of detecting diverse cardiac abnormalities and diseases using few training samples. Methods We utilized publicly available datasets like PTB-XL, CPSC, etc. To build our model, totaling approximately 80,000 ECGs. These 12 lead ECGs are scaled and filtered before using them for training. The AI model is pretrained using Self Supervised Learning and Contrastive Learning strategies. Further, the model undergoes supervised training using a hierarchy of labels that increase in granularity, encompassing arrhythmias, heart blocks, other cardiac abnormalities, heart failure, sudden cardiac death and more. Lastly, the model is prepared for Few-Shot Learning — where the model learns from a minimal set of samples (5 to 10 ECGs) of the target disease. For testing the model, we are using unique datasets like PROSE which consists of non-ischemic cardiomyopathy patients with and without myocardial scaring detected using cardiac MRI imaging at the Johns Hopkins University. Results Our model, trained on a dataset of only 5 ECG recordings, demonstrated a 74% accuracy in identifying myocardial scarring as detected by LGE MRI, utilizing a standard 12-lead ECG. This performance is comparable to that of conventional deep learning methods, which necessitate datasets containing upwards of 120 ECG recordings. Conclusion Our model represents a promising stride in creating a unified and adaptable foundation AI model for comprehensive ECG-based diagnosis. Its capacity to learn from a minimal set of examples signifies a robust framework that can evolve and adapt to new challenges in ECG diagnosis. Ongoing improvements, particularly through dataset expansion, underscore its potential to enhance cardiovascular diagnostics, offering precise and efficient patient care.

Discovery of Plasma Lipids as Potential Biomarkers Distinguishing Breast Cancer Patients from Healthy Controls

The development of a sensitive and specific blood test for the early detection of breast cancer is crucial to improve screening and patient outcomes. Existing methods, such as mammography, have limitations, necessitating the exploration of alternative approaches, including circulating factors. Using 598 prospectively collected blood samples, a multivariate plasma-derived lipid biomarker signature was developed that can distinguish healthy control individuals from those with breast cancer. Liquid chromatography with high-resolution and tandem mass spectrometry (LC-MS/MS) was employed to identify lipids for both extracellular vesicle-derived and plasma-derived signatures. For each dataset, we identified a signature of 20 lipids using a robust, statistically rigorous feature selection algorithm based on random forest feature importance applied to cross-validated training samples. Using an ensemble of machine learning models, the plasma 20-lipid signature generated an area under the curve (AUC) of 0.95, sensitivity of 0.91, and specificity of 0.79. The results from this study indicate that lipids extracted from plasma can be used as target analytes in the development of assays to detect the presence of early-stage breast cancer.

Machine-learning models to predict the risk of recurrent VT following catheter ablation in patients with structural heart disease

Abstract Background Ventricular tachycardia (VT) is a life-threatening arrhythmia that can complicate structural heart disease. While catheter ablation is effective in treating this condition, the recurrence of VT after the procedure continues to be a concern. Consequently, there is a need for a reliable system to evaluate the likelihood of arrhythmia recurrence following an ablation. Objective To predict recurrence or repeat ablation after a VT catheter ablation procedure using machine learning (ML) models. Methods We conducted a single-center, retrospective study of all patients undergoing catheter ablation for scar-related VT between 2012 and 2022. Data collected included demographics, comorbidities, medications, relevant laboratory abnormalities, electrocardiograms, echocardiograms, detailed procedural characteristics, and outcomes. Python version 3.8.5 for exploratory data analysis (EDA) and visualization. Using six different ML models from the training set (90:10 split), we used 34 variables to predict the primary outcome, including VT recurrence or repeat catheter ablation. The accuracy of these models was compared using ROC curves based on their performance on the test set. Results Out of 508 VT ablation procedures, 261 experienced a recurrence. Support Vector Classifier (SVM) performed the best, having an AUC of 0.73, followed by logistic regression at 0.69 (Figure A). SVM yielded sensitivity, specificity, positive and negative predictive values of 0.73, 0.73, 0.79, and 0.67, respectively. The F1 score was 0.70. Among the factors with the highest feature importance in the model, the most influential variable was the proceduralist’s experience level, followed by QRS width at baseline, LVEF, and patient’s age (Figure B). Conclusion Our model can predict recurrent VT after ablation with reasonable accuracy, outperforming the existing clinical risk scores significantly. Larger training sample sizes will help achieve more robust ML models and improve predictive accuracy further.Figure A: ROC curve depicting the AUROCB. Feature importance graph

MW‐SAM:Mangrove wetland remote sensing image segmentation network based on segment anything model

AbstractMangrove wetlands are important ecosystems in tropical and subtropical coastal areas, providing wind and wave attenuation and embankment protection functions. However, mangrove wetlands worldwide are facing severe loss and degradation. Accurate identification of mangrove wetland extent is crucial for their protection, but traditional methods struggle to meet the requirements of large‐scale, high‐precision identification. Furthermore, field data collection and annotation in wetlands in complex intertidal zones pose challenges, and the lack of high‐quality training samples limits the application of deep learning methods in this field. This paper proposes a novel semantic segmentation framework for mangrove wetlands called mangrove wetland remote sensing image segmentation network based on segment anything model (MW‐SAM) to address these issues. MW‐SAM is based on the pre‐trained SAM and achieves cross‐domain adaptation through parameter‐efficient fine‐tuning techniques. It introduces wetland‐specific prompts, auxiliary branches, semi‐supervised training, and iterative optimization strategies to improve accuracy and tackle the problem of sample scarcity. Moreover, on the mangrove wetland dataset constructed in this paper, MW‐SAM significantly outperforms SAM and other traditional methods. MW‐SAM provides a new technology for monitoring and protecting mangrove wetlands, and the constructed dataset will be made publicly available, which is expected to promote the development of mangrove wetland conservation efforts.

MTGGAN: generate type-controllable magnetic tile defect data to improve defect detection performance

ABSTRACT The scarcity of training samples and label information poses a major challenge for deep learning defect detection technology. Using generative models to expand sample data is considered an effective approach to address this issue. However, existing GAN models struggle to generate high-fidelity and type-controllable magnetic tile defect samples. Therefore, this paper proposes a type-controllable magnetic tile generation method (MTGGAN) that generates data with inherent label information. The model enhances fidelity and diversity in conditional generation through dual discriminator, auxiliary classifier, and auxiliary generator mechanisms. Additionally, focal loss is introduced to address the model’s class tendency issue during training. Extensive experiments demonstrate that, compared to advanced GAN methods, samples generated by MTGGAN exhibit superior quality and diversity. Furthermore, the generated data can train superior detection models. Effectively alleviating the problem of data scarcity in defect detection.

Machine Learning techniques for predicting inpatient probability and LOS from EU Injury Database

Abstract Background Unlike statistical inference Machine Learning (ML) makes repeatable predictions without prior assumptions about the underlying relationships among variables. The Full Data Set (FDS) of the EU-IDB (European Injury DataBase) consent to explore risk of hospitalization due to injuries using many predictors. The Length Of Stay of patients in hospital (LOS) is a relevant proxy of complexity of treatment and resources consumption. Methods The IDB-FDS provides more than 3.800.000 ED records, in years 2008-19 for 19 Countries. LASSO (Least Absolute Shrinkage and Selection Operator) cross-validated linearized regression technique (linear for LOS) was used for variable selection and parameter regularization. Inpatients are those admitted or transferred to hospital. Days of hospitalization were used for LOS. A cross-validated Generalized Linear Model was performed on 5 folds randomly sampled assigning 80% of records to training and 20% to testing samples. Results The strongest predictors of hospital admission risk, selected by the model were in order of importance: EUROCOST-39 diagnoses categories, Age Group, Intent, Mechanism Of Injury, Activity When Injured, Transport Injury Event, Sex Of Patient, Place Of Occurrence. EUROCOST-39 represents 61,9% of explained variability and age group 19,4%. The strongest predictors of LOS were substantially the same: EUROCOST-39 86,2% of explained variability and age group 8,6%. Conclusions The main part of variability in the ML model is explained by diagnoses reclassified according to a disability standardization method. For instance, in the maximum training sample risk of hospitalization ranges from odd 0,76% for hand/fingers sprain up to 154,02% for brain concussion. In the median sample LOS ranges from 0,45 days for strain of hand/fingers up to 11,65 for multi-trauma. A combination of more disabling injury, older age and mechanism of injury (i.e. threat to breathing) increases enormously the risk of hospitalization and LOS. Key messages • Machine Learning techniques applied on EU-IDB can provide identification of relevant risk factors of hospital admission from injuries for targeting preventive measures and organizing health care. • Artificial Intelligence consents to analyse big amount of data as dimension or analytical detail of information for identifying patterns and predictors of specific indicators for injuries or diseases.

Clinically applicable optimized periprosthetic joint infection diagnosis via AI based pathology

Periprosthetic joint infection (PJI) is a severe complication after joint replacement surgery that demands precise diagnosis for effective treatment. We enhanced PJI diagnostic accuracy through three steps: (1) developing a self-supervised PJI model with DINO v2 to create a large dataset; (2) comparing multiple intelligent models to identify the best one; and (3) using the optimal model for visual analysis to refine diagnostic practices. The self-supervised model generated 27,724 training samples and achieved a perfect AUC of 1, indicating flawless case differentiation. EfficientNet v2-S outperformed CAMEL2 at the image level, while CAMEL2 was superior at the patient level. By using the weakly supervised PJI model to adjust diagnostic criteria, we reduced the required high-power field diagnoses per slide from five to three. These findings demonstrate AI’s potential to improve the accuracy and standardization of PJI pathology and have significant implications for infectious disease diagnostics.

Prediction of optical chaos using a multi-stage extreme learning machine with data uncertainty

In this paper, we study the problem of predicting optical chaos for semiconductor lasers, where data uncertainty can severely degrade the performance of chaos prediction. We hereby propose a multi-stage extreme learning machine (MSELM) based approach for the continuous prediction of optical chaos, which handles data uncertainty effectively. Rather than relying on pilot signals for conventional reservoir learning, the proposed approach enables the use of predicted optical intensity as virtual training samples for the MSELM model learning, which leads to enhanced prediction performance and low overhead. To address the data uncertainty in virtual training, total least square (TLS) is employed for the update of the proposed MSELM’s parameters with simple updating rule and low complexity. Simulation results demonstrate that the proposed MSELM can execute the continuous optical chaos predictions effectively. The chaotic time series can be continuously predicted for a time period in excess of 4 ns with a normalized mean squared error (NMSE) lower than 0.012. It also demands much fewer training samples than state-of-the-art learning-based methods. In addition, the simulation results show that with the help of TLS, the length of prediction is improved significantly as the uncertainty is handled properly. Finally, we verify the prediction ability of the multi-stage ELM under various laser parameters, and make the median boxplot of the predicted results, which shows that the proposed MSELM continues to produce accurate and continuous predictions on time-varying optical chaos.

Enhanced Quantum Entanglement Detection of General Two Qubits Systems Based on Modified CNN‐BiLSTM Model

AbstractEntanglement is a key element in quantum information processing. The detection of entanglement is crucial in many long‐range quantum information tasks, including secure communication and fundamental tests of quantum physics, but it is also highly resource‐intensive. Even for simple 2‐qubits systems, satisfactory detection is challenging. In this work, a modified entanglement detection model combining a convolutional neural network (CNN) and a bidirectional long short‐term memory network (BiLSTM) is proposed. It shows that the proposed model can effectively extract the deep features and correlations, enabling accurate classification of simple quantum states, even with only a few tens of training samples. When trained with a large number of highly random samples, the model exhibits outstanding fitting capability, resulting in the reliable classification of nearly all common 2‐qubits systems. Furthermore, the model exhibits exceptional adaptability and significant application potential in higher‐dimensional systems.

Zero day ransomware detection with Pulse: Function classification with Transformer models and assembly language

Finding automated AI techniques to proactively defend against malware has become increasingly critical. The ability of an AI model to correctly classify novel malware is dependent on the quality of the features it is trained with and the authenticity of the features is dependent on the analysis tool. Peekaboo, a Dynamic Binary Instrumentation tool defeats evasive malware to capture its genuine behaviour. The ransomware Assembly instructions captured by Peekaboo, follow Zipf’s law, a principle also observed in natural languages, indicating Transformer models are particularly well-suited to binary classification. We propose Pulse, a novel framework for zero day ransomware detection with Transformer models and Assembly language. Pulse, trained with the Peekaboo ransomware and benign software data, uniquely identify truly new samples with high accuracy. Pulse eliminates any familiar functionality across the test and training samples, forcing the Transformer model to detect malicious behaviour based solely on context and novel Assembly instruction combinations.

Diagnosis of invasive encapsulated follicular variant papillary thyroid carcinoma by protein-based machine learning.

Although the criteria for follicular-pattern thyroid tumors are well-established, diagnosing these lesions remains challenging in some cases. In the recent World Health Organization Classification of Endocrine and Neuroendocrine Tumors (5th edition), the invasive encapsulated follicular variant of papillary thyroid carcinoma was reclassified as its own entity. It is crucial to differentiate this variant of papillary thyroid carcinoma from low-risk follicular pattern tumors due to their shared morphological characteristics. Proteomics holds significant promise for detecting and quantifying protein biomarkers. We investigated the potential value of a protein biomarker panel defined by machine learning for identifying the invasive encapsulated follicular variant of papillary thyroid carcinoma, initially using formalin- fixed paraffin-embedded samples. We developed a supervised machine-learning model and tested its performance using proteomics data from 46 thyroid tissue samples. We applied a random forest classifier utilizing five protein biomarkers (ZEB1, NUP98, C2C2L, NPAP1, and KCNJ3). This classifier achieved areas under the curve (AUCs) of 1.00 and accuracy rates of 1.00 in training samples for distinguishing the invasive encapsulated follicular variant of papillary thyroid carcinoma from non-malignant samples. Additionally, we analyzed the performance of single-protein/gene receiver operating characteristic in differentiating the invasive encapsulated follicular variant of papillary thyroid carcinoma from others within The Cancer Genome Atlas projects, which yielded an AUC > 0.5. We demonstrated that integration of high-throughput proteomics with machine learning can effectively differentiate the invasive encapsulated follicular variant of papillary thyroid carcinoma from other follicular pattern thyroid tumors.

Enhancing Few-Shot Out-of-Distribution Detection with Pre-Trained Model Features.

Ensuring the reliability of open-world intelligent systems heavily relies on effective out-of-distribution (OOD) detection. Despite notable successes in existing OOD detection methods, their performance in scenarios with limited training samples is still suboptimal. Therefore, we first construct a comprehensive few-shot OOD detection benchmark in this paper. Remarkably, our investigation reveals that Parameter-Efficient Fine-Tuning (PEFT) techniques, such as visual prompt tuning and visual adapter tuning, outperform traditional methods like fully fine-tuning and linear probing tuning in few-shot OOD detection. Considering that some valuable information from the pre-trained model, which is conducive to OOD detection, may be lost during the fine-tuning process, we reutilize features from the pre-trained models to mitigate this issue. Specifically, we first propose a training-free approach, termed uncertainty score ensemble (USE). This method integrates feature-matching scores to enhance existing OOD detection methods, significantly narrowing the gap between traditional fine-tuning and PEFT techniques. However, due to its training-free property, this method is unable to improve in-distribution accuracy. To this end, we further propose a method called Domain-Specific and General Knowledge Fusion (DSGF) to improve few-shot OOD detection performance and ID accuracy under different fine-tuning paradigms. Experiment results demonstrate that DSGF enhances few-shot OOD detection across different fine-tuning strategies, shot settings, and OOD detection methods. We believe our work can provide the research community with a novel path to leveraging large-scale visual pre-trained models for addressing FS-OOD detection. The code will be released.

OmicsFootPrint: a framework to integrate and interpret multi-omics data using circular images and deep neural networks.

The OmicsFootPrint framework addresses the need for advanced multi-omics data analysis methodologies by transforming data into intuitive two-dimensional circular images and facilitating the interpretation of complex diseases. Utilizing deep neural networks and incorporating the SHapley Additive exPlanations algorithm, the framework enhances model interpretability. Tested with The Cancer Genome Atlas data, OmicsFootPrint effectively classified lung and breast cancer subtypes, achieving high area under the curve (AUC) scores-0.98±0.02 for lung cancer subtype differentiation and 0.83±0.07 for breast cancer PAM50 subtypes, and successfully distinguished between invasive lobular and ductal carcinomas in breast cancer, showcasing its robustness. It also demonstrated notable performance in predicting drug responses in cancer cell lines, with a median AUC of 0.74, surpassing nine existing methods. Furthermore, its effectiveness persists even with reduced training sample sizes. OmicsFootPrint marks an enhancement in multi-omics research, offering a novel, efficient and interpretable approach that contributes to a deeper understanding of disease mechanisms.

Deep learning assisted cancer disease prediction from gene expression data using WT-GAN

Several diverse fields including the healthcare system and drug development sectors have benefited immensely through the adoption of deep learning (DL), which is a subset of artificial intelligence (AI) and machine learning (ML). Cancer makes up a significant percentage of the illnesses that cause early human mortality across the globe, and this situation is likely to rise in the coming years, especially when non-communicable illnesses are not considered. As a result, cancer patients would greatly benefit from precise and timely diagnosis and prediction. Deep learning (DL) has become a common technique in healthcare due to the abundance of computational power. Gene expression datasets are frequently used in major DL-based applications for illness detection, notably in cancer therapy. The quantity of medical data, on the other hand, is often insufficient to fulfill deep learning requirements. Microarray gene expression datasets are used for training procedures despite their extreme dimensionality, limited volume of data samples, and sparsely available information. Data augmentation is commonly used to expand the training sample size for gene data. The Wasserstein Tabular Generative Adversarial Network (WT-GAN) model is used for the data augmentation process for generating synthetic data in this proposed work. The correlation-based feature selection technique selects the most relevant characteristics based on threshold values. Deep FNN and ML algorithms train and classify the gene expression samples. The augmented data give better classification results (> 97%) when using WT-GAN for cancer diagnosis.

Machine Learning-Driven Data Valuation for Optimizing High-Throughput Screening Pipelines.

In the rapidly evolving field of drug discovery, high-throughput screening (HTS) is essential for identifying bioactive compounds. This study introduces a novel application of data valuation, a concept for evaluating the importance of data points based on their impact, to enhance drug discovery pipelines. Our approach improves active learning for compound library screening, robustly identifies true and false positives in HTS data, and identifies important inactive samples in an imbalanced HTS training, all while accounting for computational efficiency. We demonstrate that importance-based methods enable more effective batch screening, reducing the need for extensive HTS. Machine learning models accurately differentiate true biological activity from assay artifacts, streamlining the drug discovery process. Additionally, importance undersampling aids in HTS data set balancing, improving machine learning performance without omitting crucial inactive samples. These advancements could significantly enhance the efficiency and accuracy of drug development.

Robot learning on the job: Human-in-the-loop autonomy and learning during deployment

With the rapid growth of computing powers and recent advances in deep learning, we have witnessed impressive demonstrations of novel robot capabilities in research settings. Nonetheless, these learning systems exhibit brittle generalization and require excessive training data for practical tasks. To harness the capabilities of state-of-the-art robot learning models while embracing their imperfections, we present Sirius, a principled framework for humans and robots to collaborate through a division of work. In this framework, partially autonomous robots are tasked with handling a major portion of decision-making where they work reliably; meanwhile, human operators monitor the process and intervene in challenging situations. Such a human–robot team ensures safe deployments in complex tasks. Further, we introduce a new learning algorithm to improve the policy’s performance on the data collected from the task executions. The core idea is re-weighing training samples with approximated human trust and optimizing the policies with weighted behavioral cloning. We evaluate Sirius in simulation and on real hardware, showing that Sirius consistently outperforms baselines over a collection of contact-rich manipulation tasks, achieving an 8% boost in simulation and 27% on real hardware than the state-of-the-art methods in policy success rate, with twice faster convergence and 85% memory size reduction. Videos and more details are available at https://ut-austin-rpl.github.io/sirius/ .